Upcoming changes (not yet released)
Changes not yet released. Updated areas: Language/System, Lattices, Matrix Groups, Quaternion Algebras, Representation Theory and 36 more.
Changes and Removals
Language/System
Fixed default Windows installations leaving
magmaunavailable on PATH. The PATH box is ticked by default for both install modes, somagmaruns in a fresh Command Prompt or PowerShell. Users can untick it. Upgrades keep the earlier choice; an upgrade from an old per-user default install needs the box ticked once. The bundled DLLs on PATH can conflict with other programs. (Issue #462)About half of the Handbook example files in the
examplesdirectory of the library root (GetLibraryRoot) have new names, with no aliases for their V2.30-1 names.AlgQEA_QGrpAutoms2.mandModAlg_IrreducibleModules2.mkeep their V2.30-1 names but now hold different examples. Each example keeps the part of its file name before the first underscore, and each file’s first line gives the path of its page in the online Handbook.Each Linux download is one archive holding the shared installation tree and the generic, AVX2 and AVX2 large engines. Extract it once, then name the engine you want
magma.exe, asINSTALL.mddescribes.
Lattices
GenusRepresentativesnow raises an error, as not implemented, for a number field lattice with a non-trivial involution thatIsDefinitedoes not call definite. Previous versions could return too few isometry classes for such a lattice unless its involution was complex conjugation on a CM field. Over a CM field with complex conjugation as the involution,IteratedNeighboursnow raises an error for an indefinite one. It returned a list of lattices there when given a finiteLimit, and otherwise could fail to terminate.IsIsometricnow raises an error for two such lattices when the first is indefinite. On those inputs, it raised an error if the second lattice was positive definite, and otherwise returnedfalse, even for a lattice and itself.
Matrix Groups
CompositionTreeNonAbelianFactorscould leave out non-abelian composition factors thatCompositionTreeCompositionFactorslisted for the same group. It now returns exactly the non-abelian entries ofCompositionTreeCompositionFactors, in the same order, as a sequence of triples in the format ofCompositionFactors. It used to return a list whose triples had strings such as"A"or"J4"in place of some integers, so code reading those entries needs updating.CompositionTreeCompositionFactorsdescribed the Tits group \({}^2F_4(2)'\) by<15, 2, 2>, which names no simple group; it now gives<15, 4, 2>. Reported by David Craven.
Quaternion Algebras
Embed(Oc, O)for a quadratic order and an Eichler order overZin an indefinite quaternion algebra now defaults to solving the ternary norm form on the trace coset rather than to conjugating an embedding of the quadratic field into the order. Both answer the same question and any embedding is as good as any other, but the element returned for a given input may differ from the one the previous default returned. Definite algebras still enumerate, and orders over a number ring are unaffected.The parameter
Algains the value"TernaryForm"for the new method, andAl := "Conjugation"still selects the previous one, including its failure to terminate on some inputs; where one of its passes makes no progress at all it now says so instead of repeating that pass forever.Al := "Auto"uses the new method and falls back to conjugating whenever the search does not reach an element, whatever stopped it, so it can inherit that failure to terminate on an input the new method does not answer.Al := "TernaryForm"reports the new method’s own outcome and never falls back. (Issue #59)
Representation Theory
GaloisRepresentationandWeilRepresentationignored theirPrecisionparameter on a supercuspidal local component with an admissible pair, and always built the representation at precision 10. They now use it andAdmissiblePairtakes it too; the default is now 12, and a smaller request is raised to 12. The precision needed cannot be predicted from the conductor, and a computation on the result, such asEpsilonFactor, can need more than the construction did.For a root datum that is not semisimple, the weights of a
LieRepresentationDecompositionare now characters of the maximal torus in standard coordinates, as inHighestWeightRepresentation. Previous versions used fundamental weight coordinates of each simple component; the two agree when the simple coroots are standard basis vectors, as for root data built withIsogeny := "SC". Otherwise, as for a G2, F4 or E8 component in its default realisation, the same weight now gives a different representation, and the rows of a restriction matrix forBranchorCollectare characters of the target. On such root data,TensorProduct,TensorPower,SymmetricPower,AlternatingPower,Plethysm,AdamsOperatorandDominantCharacteralso raised errors about non-dominant weights, andSpectrumreturned wrong multiplicities; these now work.ToLiEandFromLiEstill use fundamental weight coordinates, soFromLiEconverts weights written for previous versions.For a root datum \(R\) that is not semisimple,
QuantumDimensionread the firstRank(R)entries of the weight as weight coordinates, and rejected negative entries. It now reads the whole weight as a character of the maximal torus in standard coordinates, asRepresentationDimensionnow does, and accepts any dominant character. Code that passed weight coordinates must pass the character instead, or it gets a different result or an error; semisimple root data are unchanged.EpsilonFactorandRootNumbernow raise a not-implemented error for a Galois representation of degree 1 with conductor exponent greater than 2, where previous versions could return a wrong value. This also applies to Galois and Artin representations whose epsilon factor depends on such a character, for example a quadratic character of conductor divisible by 8.
New Features
Algebraic Geometry
PointsModPN,PointsandRationalPointsnow return the points of an affine scheme over \(O/M\), where \(O\) is a maximal order of a number field and the ideal \(M\) has a non-principal prime factor. Before, such primes raised an error; they still do when \(O\) is a non-maximal order or an order of a function field.
Commutative Algebra
MinimalDecompositionhas a new formMinimalDecomposition([RngMPol], [RngMPol])for the sequences \(Q\) of primary components and \(P\) of associated primes returned byPrimaryDecomposition. It returns a minimal decomposition \(M\) of \(\cap_{J \in Q} J\) taken from \(Q\) and the sequence \(E\) of the matching primes from \(P\). Its third return value is the sequence \(N\) of integers with \(M[j] = Q[N[j]]\) and \(E[j] = P[N[j]]\).The form
MinimalDecomposition([RngMPol])now also returns such a sequence \(N\), with \(M[j] = S[N[j]]\) for the input sequence \(S\). It no longer accepts a set of ideals, only a sequence.
Groebner Bases
The Faugere \(F_4\) algorithm used by
GroebnerBasisand related intrinsics overGF(p)for odd primes \(p < 2^{63}\) or \(2^{64} < p < 2^{127}\) is faster than in previous versions whenSetNthreadsis above 1 (the speedup is typically a factor of 1.5 to 1.8). Results are unchanged for all inputs on which the previous version was correct. With many threads memory use is lower than before (with a few threads it is slightly higher for primes of about 29 to 63 bits).
Language/System
The macOS disk image now shows an Applications link, so you can drag the Magma app onto it to copy the app to
/Applications. Reported by John Voight and Don Taylor.
Modular Forms
WeilRepresentation(RepLoc)andGaloisRepresentation(RepLoc)now construct the representation attached to a supercuspidal local component of \(GL_2(\mathbb{Q}_2)\) of conductor exponent 3 when the newform has trivial character. Such a component is exceptional: it comes from no admissible pair, and earlier releases refused it for that reason.Supercuspidal components of conductor exponent 3 whose newform has nontrivial character are still refused, and exceptional supercuspidals of higher conductor exponent remain unsupported.
Optimization
Copynow provides an independent copy of an LP process, preserving its constraints, objective function, bounds, integer-variable flags, optimisation direction, and coefficient and solution rings. Assignment shares the process; this is documented besideLPProcessand the operations that change it. Reported by Markus Grassl.
Quadratic Forms
New intrinsic
ReductionOrbitMinimum(f) for a binary quadratic form f of positive discriminant: a canonical representative of the proper equivalence class of f, namely the least of the reduced forms equivalent to f in the lexicographic order on their coefficients [a, b, c]. It walks the orbit of reduced forms without storing it, so it is much faster than taking the least form ofReductionOrbit(f) when the orbit is long.
Rational Curves and Conics
Automorphism(CrvCon, AlgQuatElt)now works for a conic over a field of characteristic 2, where it used to refuse. As in odd characteristic,Automorphism(C, a*b)isAutomorphism(C, a)*Automorphism(C, b), the map that appliesAutomorphism(C, a)first. Reported by Eran Assaf.QuaternionAlgebra(CrvCon)andIsIsomorphic(CrvCon, CrvCon)raised an error for every conic over a field of characteristic 2.QuaternionAlgebra(CrvCon)now works over any field of characteristic 2, andIsIsomorphic(CrvCon, CrvCon)over a finite one. In characteristic 2,IsIsomorphic(CrvCon, CrvCon)still raises an error over an infinite field. Reported by Eran Assaf.
Real and Complex Fields
Rootsfinds the complex roots of a polynomial over the integers, the rationals, a real field or a complex field faster than in previous versions. Some polynomials of degree a few hundred, at a few hundred digits, took minutes and now take about a second.Clusters of nearly equal roots now typically come back to the precision asked for, where previous versions could return far fewer correct digits. This holds by default and with
Al := "Schonhage", and inRoots(f, C)whenfis over a complex field more precise thanC. In rare cases no method can guarantee the precision asked for, and the roots returned may then fall short of it without warning.
Representation Theory
EpsilonFactorandRootNumbernow return a value for a wildly ramified Galois representation of degree 1 with conductor exponent 2 whose character has order \(pt\), \(t\) prime to \(p\). They raised a not-implemented error for these unlessGroup(GalRep)of the representation was cyclic of order \(p\). Artin representations with such a character as a local component, and some representations induced from one, including some thatGaloisRepresentationreturns for elliptic curves, now answer too. The error remains when \(t/\gcd(t, d+2) > 2\), \(d\) being the exponent of the different of the base field over \(\mathbb{Q}_p\), or when Magma cannot construct the field cut out by the character.The Artin conductor of a relative Artin representation is now available over its own base field \(F\), rather than only as the rational conductor of the induction to \(\Q\):
ConductorIdeal(A)returns it as an ideal of the ring of integers of \(F\), andConductorExponent(A, P)its exponent at a prime \(P\) of \(F\). The two are related to the existingConductor(A)by conductor-inductivity, \(\mathrm{Conductor}(A) = |d_F|^{\dim A} \cdot N_{F/\Q}(\mathrm{ConductorIdeal}(A))\), with \(d_F\) the absolute discriminant of \(F\) and \(N_{F/\Q}\) the absolute norm. Representations whose field is not its own Galois closure are refused with an error, andConductorIdealalso refuses when the discriminant of the defining polynomial of that field is not completely factored and its unfactored part has a factor in common with the discriminant of the field.Local data of a relative Artin representation is now reachable at a prime \(P\) of its base field \(F\), which may be given as a tower:
DecompositionGroup(A, P),InertiaGroup(A, P)andRamificationGroup(A, P, i)return the corresponding subgroups ofGroup(A), which is \(\mathrm{Gal}(L/F)\) for \(L\) the Galois closure ofField(A), andLocalCharacter(A, P)the restriction of \(A\) to the decomposition group there. The groups are those of one prime ofField(A)above \(P\), so they are determined by \(P\) up to conjugacy inGroup(A). Ramification groups use lower numbering, in which \(G_{-1}\) is the decomposition group and \(G_0\) the inertia group. Representations whose field is not its own Galois closure are refused with an error.EulerFactornow accepts a prime ideal \(P\) of \(F\) for an Artin representation \(A\) returned byArtinRepresentations(E), where \(E\) is an extension of a number field \(F \ne \mathbb{Q}\).EulerFactor(A, P)is the local factor of \(A\) at \(P\), ramified or not, as a polynomial in \(T = p^{-s}\) for the rational prime \(p\) below \(P\).EulerFactor(A, p)still returns the Euler factor of the induction of \(A\) to \(\mathbb{Q}\), which is the product of these factors over the primes of \(F\) above \(p\).EulerFactor(A, P)reports an error when \(E\) is not Galois over \(\mathbb{Q}\).IsRamifiedandIsWildlyRamified, which reported an error for these representations, now answer for the induction of \(A\) to \(\mathbb{Q}\), asEulerFactor(A, p)does. SoIsRamified(A, p)istrueat every prime \(p\) that ramifies in \(F\), even when \(A\) is unramified at all primes of \(F\) above \(p\).WeylDimensionhas a new optionalBasisparameter, accepting"Weight","Root", or"Standard".
Bug Fixes
Abelian Extensions
Fixed
DecompositionGroup,DecompositionFieldandDecompositionTypefor an abelian extension at an infinite place where it is unramified. At a place outside the modulus, the decomposition group was the whole group and the decomposition field was the base field. At a place in the modulus, the decomposition type said the place ramifies. Now the decomposition group is trivial there, and every entry of the decomposition type is<1, 1>. (Issue #173) Reported by Edgar Costa.For abelian extensions with non identical but equal coefficient rings
eq,subset,meetand*have been fixed. (Issue #182)Fixed
Automorphismson abelian extensions returning a map structure instead of a sequence of relative automorphisms. (Issue #190) Reported by Edgar Costa.DecompositionGroupfor finite places and abelian extensions has been fixed. (Issue #191)Fixed an error in
AutomorphismGroupwithAllfor abelian extensions of degree one. It returns the automorphisms of the base field. (Issue #448) Reported by Edgar Costa.Fixed an error in
AutomorphismGroupfor degree-one abelian extensions whenOveris supplied explicitly. (Issue #448) Reported by Edgar Costa.Fixed an error in
ProbableAutomorphismGroupfor class fields of degree one over their base field. It returns the base field’s automorphism group and a single candidate. (Issue #448) Reported by Edgar Costa.AutomorphismGroup(FldAb)withAllandGensnow always returns a sequence of automorphisms that generate the group. Before, it ignoredGensand returned the group when an earlier call on the same extension, such asAutomorphismGroup(A : All)orTwoCocycle, had already computed it. (Issue #454) Reported by Edgar Costa.
Aggregates
Comparing tuples with
eqornecould give a wrong answer, or crash, when a component was itself a tuple. Inner tuples are now compared as outer ones are: incompatible components give the error “Arguments are not compatible”, and compatible ones are compared after coercion. (Issue #458) Reported by Eran Assaf.A set written as a range, such as
{a..b by k}, could compare unequal to the same set given otherwise when it had an element of absolute value at least \(2^{30}\) (in the Standard version; \(2^{62}\) in the Large version), a step \(k\) at least that large, or a single element and a step other than 1, and it could then have the wrong hash (for example in a set of sets); comparing such a set with a large stepkcould crash. This has been fixed. (Issue #457) Reported by Eran Assaf.Copying an associative array whose keys are types (for example, by passing it to a function which assigns to it), or creating an associative array with the universe of such an array, gave an error. This has been fixed. Reported by Edgar Costa.
Algebraic Curves
IsIsomorphicraised an error in V2.30-1 on two distinct genus 3 plane quartics over a field of characteristic 0 or greater than 7 when either curve was affine. It also raised an error over any field when a genus 3 curve in \(\mathbb{P}^1 \times \mathbb{P}^1\) was compared with itself, with another such curve or with a plane quartic. It now returnsfalsewhen such curves are not birationally equivalent, and for curves in \(\mathbb{P}^1 \times \mathbb{P}^1\) it returnstrueand a birational map between them when they are. In characteristic 2, 3, 5 or 7, it previously raised an error too when an affine genus 3 plane quartic was compared with an equal curve. It now returnstrueand the identity map for such a pair. Two distinct birationally equivalent curves still raise an error when either of them is affine. (Issue #238)
Algebraic Function Fields
Automorphismsof a function field now results in a runtime error instead of an assertion fail when theAutomorphismsof the constant field cannot be computed. (Issue #214) Reported by Adam Logan.
Algebraic Geometry
Fixed incorrect integer lifts of p-adic coefficients in
EulerFactorfor Jacobi motives with more than 16 Frobenius roots. Reported by John Voight.
Algebraic Number Fields
ResidueClassFields of ideals of orders of algebraic number fields are now handled more similarly when the Montes algorithm is used for prime splitting to when the Montes algorithm isn’t used. Reported by Raymond van Bommel.CyclotomicUnitson a cyclotomic field of degree 1,CyclotomicField(1)orCyclotomicField(2), corrupted memory and could crash Magma later in the session. It now returns the same empty sequence without the corruption.A memory issue whilst coercing places into divisors has been fixed. (Issue #172)
SUnitDiscLoghas been fixed when the second argument is an element rather than a sequence. (Issue #187)IsMaximalfor orders whose coefficients rings are not yet known to be maximal has been fixed. (Issue #194)IsIsomorphicandIsSubfieldstopped with an internal error onFieldOfFractions(OA)andAbsoluteField(N), forNa nonsimple relative field. HereOAis the absolute maximal order ofN, built before the absolute field. Both now returntrue, andIsIsomorphicalso returns an isomorphism. (Issue #196) Reported by Edgar Costa.Over a field built with
QuadraticField, these stopped with an assertion failure:SupportandCoprimeBasisof a sequence of ideals,Supportof a sequence of field elements, andCoprimeBasisInsert. They now give the same results as over the same field built withNumberField. (Issue #204) Reported by Edgar Costa.PrincipalMultiplehas been fixed for ideals of non maximal orders. (Issue #221)ClassGroupcould crash on the maximal order of a subfield obtained withFixedFieldandRelativeFieldfrom the field of fractions of an absolute order. Whether it crashed depended on the random seed. It no longer crashes there. (Issue #222) Reported by Edgar Costa.The precision of
RootNumbergiven a Grossencharacter or a Hecke character and an integer has been fixed.Applying the map from the
PicardGroupnow results in an integral ideal coprime to the conductor which can be passed in to the preimage of the map. (Issue #251) Reported by Stefano Marseglia.CoprimeBasisof a sequence of ideals of an order of a number field has been fixed for ideals of relative extensions. (Issue #440)Decompositionof a prime in the maximal order of a tower of three or more relative extensions of number fields had a memory management problem, which could cause crashes. This has been fixed. Reported by Eran Assaf.Fixed an error when inverting ideals of a maximal order returned by
OptimizedRepresentationwithAl := "Fast"orAl := "Polredabs". (Issue #436) Reported by Edgar Costa.Fixed
SimNEQfailing for relative number fields, including calls with supplied prime ideals. (Issue #443) Reported by Edgar Costa.Fixed errors in
UnitsWithSignswhen the number of requested real places differs from the number of unit group generators, including requests with no real places.HasTotallyPositiveGenerator, which calls it, failed in the same way on fields that are not totally real. (Issue #437) Reported by Edgar Costa.
Algebraic Surfaces
HasOnlySimpleSingularitiesandIsSimpleSurfaceSingularitycould report the wrong type (e.g. A3 instead of A4) for a singular point of a surface of codimension at least 2. This has been fixed. Reported by Janez Bernik.KodairaConfigurationsnow checks that its input is an integral symmetric intersection matrix with diagonal entries -2 and nonnegative off-diagonal entries, and explains that an intersection matrix of irreducible (-2)-curves is required when these conditions fail. (Issue #461) Reported by Edgar Costa.
Algebras
Products
x*yand powersx^nof classes in the ideal class monoid returned byIdealClassMonoidAbstractcould be the wrong class, with no error raised. They now give the class of the product or power of the representative ideals. Only orders with an overorder whose Picard group is nontrivial could be affected. (Issue #171) Reported by Edgar Costa.
Arithmetic Fuchsian Groups
ShimuraReduceUnitstopped with an error before a call toGroup: “Attribute ‘ShimFDDisc’ for this structure is valid but not assigned” when no fundamental domain had been computed, and the same error for ‘ShimGroupSidepairs’ afterFundamentalDomain. AfterGroup, computing a fundamental domain in a second unit disc with another centre madeShimuraReduceUnitfail an internal assertion or not return.HeckeOperatoron Hilbert modular forms computed with that group did not return either. This has been fixed.Group,ShimuraReduceUnitby default, andFundamentalDomainwith no disc now use the group’s first fundamental domain, also afterFundamentalDomainis called with another disc.FundamentalDomainwith no disc returned the latest domain. (Issue #243) Reported by John Voight.
Coding Theory
MinimumLeeWeight(equivalentlyMinimumLeeDistance) of a quasi-cyclic code over \(\mathrm{Z}_4\) could return a value larger than the true minimum Lee weight, or \(-1\). This has been fixed. Reported by Tushar Bag.Fixed incorrect results and crashes when iterating over quantum codes with a linear normaliser. Reported by Tushar Bag and Markus Grassl.
NameandAssignNameson a linear code gave an error. This has been fixed.Fixed incorrect minimum distances for quantum codes over GF(q^2) whose stabilizer is additive over a proper subfield of GF(q). Reported by Markus Grassl.
Elliptic Curves
MordellWeilGroup(CrvEll[FldFunRat])withAl := "Descent", for an elliptic curve over a rational function field with full rational 2-torsion, printed a trace of its point search and a line with the rank at default verbosity. It now prints nothing, except the warning when it finds only a lower bound on the rank.SetVerbose("MordellWeilGroup", 1)shows the rank, and level 2 the search trace. (Issue #250) Reported by Edgar Costa.Fixed
HeegnerIndexreturning an incorrect value for a nonminimal model of an elliptic curve over the rationals. (Issue #264) Reported by Edgar Costa.Heightof a point of order 1, 2 or 3 over a number field returned zero at the default real precision, ignoringPrecision.HeightPairingandHeightPairingMatrixlikewise lost precision for points whose sum has order 1, 2 or 3, such as a point and its negative. All three now respectPrecision. (Issue #275) Reported by Edgar Costa.Fixed the inverse of the map returned by
ThreeSelmerGroupfor elliptic curves over the rationals with irreducible 3-division polynomial, which previously failed with an argument-type error inCoordinates. (Issue #285) Reported by Edgar Costa.SaturationraisedRuntime error in 'Universe': Illegal null sequencefor an elliptic curve with nonintegral coefficients or over a relative number field when given no points or only torsion points. It now returns a sequence of points of the curve: generators of the torsion subgroup, or none whenTorsionFreeistrue. (Issue #290) Reported by Edgar Costa.Fixed
LocalHeightat infinite places over relative number fields failing afterHeighthad been called on a point of the same curve. (Issue #277) Reported by Edgar Costa.Fixed an assertion failure in
IsIsogenousfor large quadratic twists of elliptic curves over the rationals with a rational isogeny of degree 11, 17, 19, 37, 43, 67 or 163. (Issue #294)The map returned by
pPowerTorsionnow maps the group identity to the elliptic identity when the returned torsion group is trivial. (Issue #293) Reported by Edgar Costa.Fixed a sequence index error in
TwoSelmerGroupandTwoDescentwith theRemoveGensparameter, which occurred on some runs depending on the random seed. (Issue #279) Reported by Edgar Costa.Added the documented
QuarticMinimizespelling for univariate and binary quartics. CorrectedQuarticMinimiseover rational function fields to return its quartic, transformation matrix and scale without a type error. The matrix uses the same convention as over the integers and the rationals. Over a rational function field,QuarticMinimisestill gives an error for a quartic with no \(x^4\) term. There it returns a cubic form for a binary quartic whose minimal quartic has no \(x^4\) term. (Issue #307) Reported by Edgar Costa.TorsionBoundandTorsionSubgroupcould raise the error “Invalid sequence universe” on an elliptic curve over \(F(t)\), \(F\) of characteristic zero, with bad reduction at more than \(n\) of \(t = 1, 2, \ldots, 2n\). Here \(n\) is the second argument ofTorsionBound, and is 7 forTorsionSubgroup. Both now return their result. (Issue #313)Fixed a runtime error in
HeegnerPointNumberOfTermswhen the naive point search found a rational point on the curve. (Issue #320)CoveringCovariantswithComputeY := falsenow returns the two requested covariants for a degree-four genus one model. Previously it gave an error unless it had already been called on that model with the defaultComputeY := true. (Issue #326) Reported by Edgar Costa.Fixed
FrobeniusMatrixfor elliptic and hyperelliptic curves over the rationals returning entries that claimed more absolute precision than computed, particularly at low requested precision or for weakly zero entries. (Issue #419) Reported by Edgar Costa.Fixed
Basisfor points on an elliptic curve over a function field, given rank and discriminant bounds. When the first point of infinite order already met both bounds, it could return extra points or a different one; it now returns that point alone. (Issue #429) Reported by Edgar Costa.Fixed
Chabautyfor finite Mordell-Weil groups to include points whose image is the rational point at infinity and to return the solutions as a set. (Issue #410) Reported by Edgar Costa.Fixed wrong results from
pAdicHeight,pAdicHeightPairingMatrixandpAdicRegulatoron models with minimal discriminant whose coefficients are \(p\)-integral but not all integers. These intrinsics now accept any model of the curve, where they raised an error unless the model had minimal discriminant and \(p\)-integral coefficients. (Issue #418) Reported by Edgar Costa.TorsionBoundfor an elliptic curve over a function field, given a third argument \(B\), could return less than the number of points of order dividing \(B\) when these form a noncyclic group. It now returns a bound on that number. (Issue #431)For twist parameters \(d\) congruent to 2 or 3 modulo 4, such as \(-1\), \(2\) and \(3\),
AnalyticRankQuadraticTwistcould give a wrong analytic rank and leading L-value, or run for a long time. It now gives those of the quadratic twist of \(E\) by \(d\), the same asAnalyticRankapplied toQuadraticTwist(E, d). Values of \(d\) congruent to 0 or 1 modulo 4 were not affected. (Issue #433)Fixed incorrect values and logarithm errors in
LocalHeightat infinite places of number fields when using low precision. (Issue #402) Reported by Edgar Costa.Fixed
EllipticCurvefailing to recognise the rational point at infinity on a genus-one hyperelliptic curve in characteristic two with a cubic right-hand side and a linear or constant coefficient of y. (Issue #330) Reported by Edgar Costa.Fixed
Height,HeightPairingandHeightPairingMatrix(on nonempty sequences) raising an error for points on elliptic curves over number fields whenPrecisionis zero. They now use the current default real precision. Positive precisions below 4 are now rounded up to 4, as forLocalHeight. (Issue #332)Fixed
AddCubicson genus one models to return both models whenReturnBothis true, instead of returning only one. (Issue #327) Reported by Edgar Costa.Fixed
TwoIsogenyDescentto reject anIsogenyparameter with the wrong source curve or degree with an input error before attempting the descent. (Issue #334) Reported by Edgar Costa.On an elliptic curve with \(j\)-invariant 0 or 1728 whose Weierstrass model has \(b_2 = a_1^2 + 4a_2 \ne 0\),
MultiplicationPolynomialcould return a wrong polynomial for some elements \(\alpha \ne \pm 1\) of the CM order. For a composite \(\alpha\),DivisionPolynomialcould also return a wrong polynomial, andCMEndomorphismcould stop with an error. All three are now right on every model; models with \(b_2 = 0\), which include every reduced minimal model over \(\mathbb{Q}\), were not affected. (Issue #420) Reported by Edgar Costa.GaloisRepresentationof an elliptic curve with wildly ramified, potentially good reduction, over a \(p\)-adic field or at a prime above \(p\), could be twisted by the wrong unramified characterUnr(c), with \(c = (j\sqrt{p})^f\) for some positive integer \(j\) and \(f\) the residue degree over \(\mathbb{Q}_p\).EpsilonFactorof the representation, the Euler factors of its tensor products, and the local factor at \(p\) ofLSeries(E, A)for an Artin representationAramified at \(p\) could therefore be wrong. Magma now twists by \(c = (\sqrt{-p})^f\).IsGenusOneModelandGenusOneModelaccepted an empty sequence, or three or more quadratic forms in four variables, as a degree-4 genus one model whose invariants ignored every form after the second. They now need exactly two forms: otherwiseIsGenusOneModelreturnsfalseandGenusOneModelraises an error.GenusOneModellikewise refuses a sequence of 4 by 4 matrices that is not a pair, and a curve in \(\mathbb{P}^3\) whose ideal is not generated by two quadrics.FourToTwoCoveringnow refuses a curve not defined by exactly two quadrics, even a genus one curve given with a redundant third equation, which it used to handle; pass itGenusOneModel(C)instead. (Issue #270) Reported by Edgar Costa.pAdicEllipticLogarithmandpAdicEllipticLogarithmOfCombinationcould return a wrong value if the session had already called either of them with the same prime andPrecision, even for a different curve. Their result no longer depends on earlier calls. (Issue #416) Reported by Edgar Costa.LocalRootNumbersfor an elliptic curve over the rationals returns the local root number at each bad prime. It previously repeated the global sign. (Issue #260)SEAwithAl := "BSGS"returned wrong point counts, or stopped with the errorFailed to find unique trace of CM curve, for many curves over fields with at most 29 elements. It also failed for some curves with \(j\)-invariant 0 or 1728 over larger fields. It now returns the same count asAl := "Default", as the handbook documents; the other settings ofAlwere not affected. (Issue #271)TwoCoverPullbackretains both signs of a quartic point above a point of order two.FourCoverPullbackreturns each point once when two-cover preimages share an x coordinate, including at infinity above a point of order two. (Issue #409) Reported by Edgar Costa.DualIsogenynow returns, for a map fromTwoIsogeny, a dual that composes with it to multiplication by two over every base field. The dual it returned before could be off by an automorphism, usually a sign, over number fields and finite fields; over the rationals it was correct. (Issue #413)
Finite Fields
Fixed an internal error in
Embedwhen embedding a default finite field after embedding subfields of incomparable degrees into the target field. (Issue #231) Reported by Edgar Costa.
Galois Groups
Fixed an assertion failure in
GaloisRootwhen supplied with a nonmonic polynomial and its Galois data. (Issue #449) Reported by Edgar Costa.
Global Fields
IdealsOfIndexfor a maximal order which is not an extension of the integers has been fixed to useAbsoluteNorm.IdealsOfIndexno longer accepts an order not known to be maximal when the index is greater than 1. The signature accepting an ideal can be used instead when the index is coprime to the conductor otherwise theAlgEtQsignatures can be used. (Issue #439)
Graphs
For a digraph, the Handbook gave 2 as the default
MaxlevelofAutomorphismGroup; it is 999, so the invariant is applied at every level of the search tree, not only the first two. The Handbook also said thatInvariant := "default"applies no invariant; it does discard an invariant remembered from an earlier call, but on a digraph the"adjacencies"invariant is still applied. The Handbook now also documents theTCLevelandVerbosityparameters, the"refinvar"invariant, every parameter thatIsIsomorphictakes for graphs, andCanonicalGraphMap. Only the documentation changed.KpartiteGraphdid not validate its argument: the empty sequence crashed Magma, a negative, oversized or non-integer part size gave an internal assertion or a misleading “too many vertices”, and a part size of 0 was accepted. These are all runtime errors now, matchingMultipartiteGraph. WithSparseRep := true,Sizeand#EdgeSetalso returned the product of the part sizes rather than the number of edges, wrong from three parts on. Reported by Reymond Akpanya.KCubeGraphaborted with an internal error onKCubeGraph(0), with or withoutSparseRep, instead of returning the \(0\)-cube \(Q_0\). It now returns the single vertex with no edges.MultipartiteGraphreported an undefined entry in its sequence as “Undefined sequence element”, and only once it had begun reading the sequence. It now refuses the sequence on arrival, with the “All of the sequence elements must be defined” the other graph constructors give. Reported by Reymond Akpanya.
Groebner Bases
On Intel Macs, threaded Groebner basis computations no longer stop with
Magma: Internal error, for exampleGroebnerBasisafterSetNthreads(2). Threaded sparse Lanczos solves are fixed the same way, for exampleModularSolutionwithLanczos := trueand a modulus of 2^62 or more. The error message named a failedsem_waitorsem_post, such assem_wait(sem_writing) failed: Bad file descriptor.GroebnerBasisand related intrinsics over GF(p) failed with an internal error whenSetNthreadswas 256 or more. This has been fixed.On arm64 builds, and on x86-64 builds made without AVX,
GroebnerBasisand related intrinsics over GF(p) for a prime p with 65537 <= p <= 11863279 could give a wrong result, an internal error or a hang whenSetNthreadswas above 1. This has been fixed.
Groups
On Windows,
BruhatDescendants, andNilpotentQuotientwith the parameterNumberOfFreeVariablesset to a positive value, could crash with a segmentation fault. Both now give the same results on Windows as on Linux. (Issue #210) Reported by Edgar Costa.
Hyperelliptic Curves
With
SetVerbose("GroupInfo", 3), the Mordell-Weil sieve for a genus 2 curve over the rationals could run forever, retrying at ever higher precision.RationalPointsGenus2runs this sieve when its search finds no point and the rank is at least 2. Raising the verbosity no longer causes this. Reported by Eran Assaf.
Incidence Geometry
IsCGroup,IsStringCGroup,IsCPlusGroup,CosetGeometryFromCGroup,CosetGeometryFromCPlusGroup,OrbitalGraphandClosureGraphcrashed on a natural symmetric group such asSym(4); they now give the same answers as for the same group given by explicit generators.
Invariant Rings
The Handbook stated the action of a matrix \(a\) on a polynomial \(f\) (as in
f ^ a) with \(a\) in place of its transpose. This has been fixed. Reported by Max Horn.
Language/System
Bug reports should now be sent to magma.bugs@sydney.edu.au, and general enquiries to magma.admin@sydney.edu.au. Both replace addresses on the University of Sydney’s retired maths.usyd.edu.au domain.
The error given for a
returnstatement outside any function did not mentionevalexpressions, and suggested that procedures may return values. This has been fixed, and the Handbook now says that areturnstatement with no expression may be used to leave a procedure early. Reported by Håvard Damm-Johnsen.Fixed
evalusing a stale comprehension loop variable after the statement containing a failed comprehension had finished. Reported by Markus Grassl.GetBinaryVarianton Windows now returns a name that ends in-windows, such asavx2-windows. Previously the Windows binary called itselfavx2-linux, also in the header of an internal error report. (Issue #207) Reported by Markus Grassl.On Windows, when HOME was unset, Magma warned that it could not create the history directory
/home/<user>/.magmaand did not keep command history between sessions. Magma uses the Windows user profile folder for history instead. (Issue #208)AttachSpec,DetachSpec,AttachandDetachon Windows now accept a path with a drive letter, such asC:/packages/spec, with forward slashes or backslashes. Previously such a path was taken as relative to the current directory, so the file was not found. (Issue #211) Reported by Edgar Costa.On Windows,
System,PipeandPOpennow run their commands: the installation includes a shell, dash, asbin/sh.exe. PreviouslyPipeandPOpenfailed andSystemdid nothing, and a failedPipeleft a second Magma reading the rest of the input. Each of the three now raises an error when the shell is missing. (Issue #212) Reported by Edgar Costa.The Start Menu and desktop shortcuts of a Windows installation for all users now start Magma in the user’s profile folder (usually
C:/Users/<name>). Previously they started in the Documents folder of the administrator who installed. (Issue #215)AssignNameson a universal enveloping algebra, and the product of a module by a module homomorphism, gave an error. This has been fixed.When
Copyrejects an unsupported object, its error message now names the object’s actual type. Reported by Markus Grassl.Opening the Magma app on macOS no longer stops at the warning “Apple could not verify ‘magma.command’ is free of malware”. The app now starts Magma in Terminal through its signed launcher. Reported by John Voight and Don Taylor.
Fixed
ChangeDirectoryfailing on Windows when~pointed to a nonexistent directory. As in POSIX shells, bare~now followsHOMEon every platform, falling back to the account’s home directory whenHOMEis unset or empty;~nameis unchanged.
Lattices
GenusRepresentativesfor a definite number field latticeL, which must have rank at least 3 if its involution is trivial, could fail to terminate when called withUseMass := falseand the default infiniteLimit. It now either returns representatives for the whole genus or raises an error asking for a finiteLimitorUseMass := true. It raises wheneverLhas a non-trivial involution. For an integral positive definiteLwith trivial involution, it returns when every prime ideal has exponent \(-1\), \(0\) or \(1\) inDiscriminant(L), and raises otherwise. For other lattices, the handbook entry gives the condition.LLLGramwithMethod := "Integral",InitialSort := trueand two or more threads could return a wrong result, or give an internal error or run out of memory. This has been fixed. (Issue #147) Reported by Edgar Costa.IsProperlyIsometricraised the error “Variable ‘f’ has not been initialized” on two number field lattices that are not isometric; it now returnsfalse. For a number field lattice,GenusRepresentativeswithProper := trueraised the same error when two lattices in its genus were not isometric but shared a theta series up toThetaPrec, which is more likely with a smallThetaPrec; it now returns the representatives. Reported by Edgar Costa.Over a CM field with complex conjugation as the involution,
IsDefinitereturnedtrue, with scaling element 1, for every hermitian lattice, including indefinite and negative definite ones. It now answers correctly, soAutomorphismGroupandGenusRepresentativeswork for a negative definite such lattice, where they raised an error.Over a totally real field,
IsDefinitereturnedtrue, with scaling element 0, for a lattice of positive rank whose Gram matrix is zero; it now returnsfalse. A lattice of rank zero over such a field is now definite, with scaling element 1, whereIsDefinitereturnedfalse. For a hermitian lattice of positive rank whose base field has a real place, which is never definite,IsDefinitecan still returntrue, though less often than before.GenusRepresentativesandSpinorRepresentativesof an indefinite binary lattice could return a lattice from another genus, miss some classes, or return isometric lattices. This has been fixed. (Issue #209) Reported by David Roe.For a norm bound above about \(10^{24}\),
CloseVectorsandNextVectoron a process fromShortVectorsProcessorCloseVectorsProcesscould use memory without limit. With coordinates of about \(10^{15}\) or more they could also skip vectors near the bound, and from \(2^{53}\) return the same vector repeatedly. Memory use is now bounded, and where coordinates would reach \(2^{53}\) they raise the errorLattice enumeration coordinates exceed floating-point precision, which a process then raises on every later call. Reported by Eran Assaf.Masscould return a wrong value or give an error for many lattices over the integers of rank 2 mod 4 and for some Hermitian lattices over imaginary quadratic fields. So couldDedekindZetaExactwithRelativefor an imaginary quadratic field at an even argument.DedekindZetaExactnow computes its values over the rationals and quadratic fields exactly. (Issue #236) Reported by David Roe.SpinorCharacterscould return wrong spinor characters, andSpinorGeneratorswrong spinor generators and so a wrong number of spinor genera, for lattices of rank at least 3: indefinite ones, and some definite ones.SpinorGeneraandIsSpinorNorm, which use them, could give wrong results too, andSpinorGeneracould raise an error. This has been fixed. (Issue #237) Reported by David Roe.For a large norm bound, short and close vector enumeration (
ShortVectorsProcess,CloseVectorsProcess,CloseVectorsand related functions) could skip vectors within the bound, as the rounding errors of its floating-point arithmetic grew with the bound; for example,ShortVectorsProcessfor the lattice with Gram matrix diag(1, 3, 3) and bound \(10^{19}\) never returned the vector (377, 35777, 1825741858) of norm \(10^{19} - 192\). It could also take practically forever to find the next vector, as for that lattice with bound \(10^{30}\) or a close vector process with a target far from the origin, or report no close vectors when there was one. When rounding could matter, the coordinate ranges are now computed exactly, so no vector is missed and none of these takes long. Reported by Eran Assaf and Allan Steel.DualBasisLatticeof a lattice of rank 0 failed with “Internal error”. It now returns the lattice of rank 0 and the same degree, asDualdoes. (Issue #393) Reported by Edgar Costa.CloseVectors,ClosestVectors,CloseVectorsMatrixandClosestVectorsMatrixfailed with the error “Variable ‘E’ has not been initialized” when both the parametersMaxandPrunewere given. This has been fixed. (Issue #388) Reported by Edgar Costa.SpinorRepresentativesgave an error for most positive definite lattices of rank at least 3, such as the standard lattice of rank 3. This has been fixed. (Issue #372) Reported by Edgar Costa.LLLwithMethod := "Integral"crashed on a zero matrix, on a matrix with no rows, and on a zero matrix withModular := true.LLLGramwithMethod := "Integral"andInitialFP := falsecrashed on a zero Gram matrix when two or more threads were used. These crashes are fixed.LLLnow returns the input with rank 0, asMethod := "FP"does. (Issue #366) Reported by Edgar Costa and Reymond Akpanya.DedekindZetaExactcould return a wrong value, or give an error, when the value was large.Massof lattices over number fields that needed such a value was then wrong or gave an error. This has been fixed. (Issue #463) Reported by Eran Assaf.SpinorRepresentativesgave an error for negative definite binary lattices, andGenusRepresentativeswithNaturalActionfor negative definite binary sublattices of an indefinite space. This has been fixed. (Issue #468) Reported by Eran Assaf.Neighboursgave an error for indefinite and negative definite binary lattices which have neighbours, except at 2 for odd lattices. For a binary lattice \(L\) of content \(c\) it could also give a wrong answer, or an error, at 2 when \(L\) is odd or \(c\) is even, and at an odd prime dividing \(\det(L)\) exactly or dividing \(c\). This has been fixed. When \(L\) is not isotropic, it still gives an error at an odd prime \(p\) such that \(p^2\) divides \(\det(L)/c^2\). (Issue #465) Reported by Eran Assaf.NeighborClosureof a binary lattice ignored the prime and returned the same lattices asGenusRepresentatives, instead of those reached by neighbours at that prime. It also gave an error for negative definite binary lattices. This has been fixed. For a binary lattice \(L\) of content \(c\) that is not isotropic, it now gives an error at an odd prime \(p\) such that \(p^2\) divides \(\det(L)/c^2\). (Issue #469) Reported by Eran Assaf.GenusRepresentativesandSpinorRepresentativesgave an error for every binary lattice representing 0 (that is, with minus its determinant a square), such as the hyperbolic plane. This has been fixed. (Issue #467) Reported by Eran Assaf.For binary lattices,
SpinorGeneratorsandSpinorGeneracould give wrong spinor generators and a wrong number of spinor genera: for indefinite lattices whose discriminant has a fundamental unit of norm +1, and for some definite lattices and lattices with content greater than 1; for proper genera,SpinorGeneracould return two lattices in the same spinor genus;SpinorGeneratorsignored the parameterProper; andSpinorGeneracould give an error for indefinite and negative definite binary lattices with more than one spinor genus. This has been fixed. (Issue #466) Reported by Eran Assaf.Neighboursignored the parameter Proper for binary lattices, returning one neighbour for each isometry class instead of each proper isometry class. This has been fixed. Reported by Eran Assaf.Neighbours,NeighborClosure,SpinorGeneratorsandSpinorGeneragave an error for binary lattices representing 0 (that is, with minus the determinant a square), such as the hyperbolic plane. This has been fixed.GenusRepresentativesandSpinorRepresentativeswithNaturalAction := truegave an error for some binary lattices, depending on the inner product of the ambient space. This has been fixed. (Issue #475) Reported by Eran Assaf.Printing a lattice of negative determinant did not show its factored determinant. It now does, with the sign before the factors (for example
-2^2). Reported by Eran Assaf.IsIsometricandIsProperlyIsometricnow compare definite quadratic lattices over a totally real field whose forms have matching signs at each real place. Previously, they could raise an error unless both forms were totally positive definite. KnownIsProperlyIsometricdefects can still give wrong answers or matrices that are not isometries, including for quadratic lattices over totally real fields. Over a CM field with complex conjugation,IsIsometricpreviously returnedfalsefor two negative definite lattices, even a lattice and itself. It now compares definite lattices whose forms have matching signs at each real place of the fixed field of the involution.IteratedNeighbourspreviously raised an error for negative definite lattices over totally real fields, or repeated isometry classes over CM fields. It now returns one lattice per isometry class. (Issue #476)
Lattices over Number Fields
/for number field lattices and ideals of orders of number fields has been fixed. (Issue #385)
Local Fields
In an extension of a local field,
Tracerelative to the coefficient field could return a wrong value at a lower precision after a trace at a higher one. The trace is now right at every precision.
Matrices
IsConsistentandSolutionwith a sequence of vectors over a ring with zero divisors, such asIntegerRing(6), could return an incomplete kernel or wrongly report a consistent system as having no solution. This has been fixed. Reported by Sergey Shpectorov.EchelonFormof a matrix over a real or complex field could return an echelon form of the wrong rank, with a transformation matrix that did not give it. This has been fixed. On an ill-conditioned matrix the rank of the echelon form can still differ fromRank. Reported by Peteris Daugulis.On Intel Macs, V2.30-1 ran linear algebra over small finite fields and Groebner bases over GF(2) up to 1.8 times slower than earlier releases. That speed is restored.
Matrix Groups
Randomon a conjugated trivial matrix group aborted the process instead of returning the identity. This has been fixed. Reported by Eran Assaf.
Modular Forms
CuspidalInducingDatumfailed an internal assertion on supercuspidal local components of \(GL_2(\mathbb{Q}_p)\) of odd conductor exponent, for \(p = 2\) and \(p \equiv 1 \pmod 4\). At \(p \equiv 1 \pmod 4\) the same assertion also stoppedAdmissiblePair,GaloisRepresentationandWeilRepresentation.CuspidalInducingDatumnow returns the cuspidal inducing datum, and those three no longer fail this assertion.Fixed an error in
Periodsfor modular symbols when the number of q-expansion terms was omitted. Reported by Edgar Costa.
Modular Symbols
AtkinLehnerOperatorandAtkinLehnerraisedRuntime error in '^': Bad argument typeson a modular symbols space with several characters and on its subspaces. They now accept such a space when every character whose summand in the ambient space is nonzero is trivial or quadratic.On a space of sign 1 or -1 with a single character, they could also return a wrong matrix for \(W_q\). That happened when such a character \(\chi\) has \(\chi(d) = -1\) for the \(d\) that is \(-1\) modulo the \(q\)-part of the level and \(1\) modulo the rest. They now raise an error there, asking for sign 0, as they do on a space with several characters, and on subspaces of either.
HeckeOperatorraised “Arguments have incompatible coefficient rings” at some primes on some subspaces of a modular symbols space of weight at least 3 with trivial or quadratic character, such as the cuspidal and new subspaces. It now returns the Hecke matrix.pAdicLSeriesreturned exact zero for the coefficients of \(T^k\) with \(k \ge p^n\), where \(n\) isCoefficientPrecision, although it had not computed them. This happened only whenSeriesPrecisionwas at least \(p^n\), so the default parameters were not affected. These coefficients now come back asO(1), and a largerCoefficientPrecisioncomputes them. WhenSeriesPrecisionis at least \(p^{n+1}\), the series now stops at \(O(T^{p^{n+1}})\) instead of at the requested length. (Issue #421) Reported by Edgar Costa.With the default
ProperScaling := true,pAdicLSeries(E, p : QuadraticTwist := D)returned 4 times the \(p\)-adic \(L\)-series of the twist of \(E\) by \(D\) when \(D\) is divisible by 8 and \(E\) has supersingular reduction at 2. Since \(p\) is odd, the valuations of the coefficients were right. It now returns the same series aspAdicLSeriesonMinimalModel(QuadraticTwist(E, D)), as it already did for other curves and discriminants. (Issue #422) Reported by Edgar Costa.
Modules
For a \(K[G]\)-module with \(K\) a number field,
CompositionFactors,ConstituentsandConstituentsWithMultiplicitiescould order their results differently from one run to the next. The order is now fixed, except that factors of equal dimension that are not submodules of the module can still change places among themselves. Reported by Eran Assaf.meet,subsetandRowspaceIsSubon subspaces over a real or complex field crashed, as didIncludeof a set of vectors. This has been fixed. Quotients of such subspaces (and soHomologywith real or complex coefficients), which also crashed, now give an error saying they are not supported. Reported by Reymond Akpanya.
Optimization
The chapter “Linear Programming” now points at the current home of the lp_solve library and lists the status codes Magma can return, as do the docstrings of
MaximalSolution,MinimalSolution,MaximalIntegerSolution,MinimalIntegerSolution,MaximalZeroOneSolutionandMinimalZeroOneSolution. Reported by Markus Grassl.
Polynomial Rings
Discriminantfor a polynomial over a quotient of a polynomial ring overIntegers(M)withMcomposite could give an internal error or a wrong result when the leading coefficient is a zero divisor. This has been fixed. Reported by Edgar Costa.
Quadratic Fields
Fixed
PellEquationraising an error when D is zero or a positive square and the target is nonzero. (Issue #452) Reported by Edgar Costa.Fixed
PellEquationraising an error for negative D and a nonzero target. (Issue #452) Reported by Edgar Costa.Fixed
PellEquationfailing to return a solution when the target is zero. (Issue #452) Reported by Edgar Costa.
Quaternion Algebras
Embedof a quadratic order into an Eichler order overZin an indefinite quaternion algebra could run forever, or stop and report that it had not found an embedding, on inputs where one exists. It searched by conjugating an arbitrary embedding of the quadratic field into the algebra one critical prime at a time, and a step taken at one prime can undo a step taken at another, so the search had no reason to reach the order. For a discriminant 210 Eichler order of level 35 and the quadratic order of discriminant -840 it ran without reaching one; how long that takes depends on the build and on what else has been computed, so on some builds it returns and on others it does not. The embedding it was looking for is143*i + 255*j + 34*k.Embednow solves for the element directly. An element of reduced tracetand reduced normnlies on the trace-tcoset of the rank 3 trace-zero sublattice, where the reduced norm restricts to a ternary quadratic form, so the question is whether that form representsnon the coset. That is decidable: the quadratic order embeds locally everywhere exactly when every prime ramified in the algebra is non-split in the quadratic field and the minimal polynomial has a root modulo each prime power dividing the Eichler level. Eichler orders of a given level in an indefinite algebra overQare all conjugate, so no selectivity arises and an order that embeds into every completion embeds globally.Embedtherefore names the local condition that fails when there is no embedding, rather than searching for one that cannot exist.Finding the element itself takes four forms, each a fallback for the one before. When the algebra is split the norm form is isotropic, and the reduced norm is linear along an isotropic direction, so a division gives the coefficient. It does not always land on an element of the order, whatever the size of
n, so this is a fast path and not a complete one. Otherwise the quadratic is solved exactly in one coordinate of a reduced basis of the trace-zero sublattice. A square factor ofnthen scales a smaller solution, becauseNrd(f*mu)isf^2*Nrd(mu)andTrd(f*mu)isf*Trd(mu); that is what keeps a largenaffordable, and it returns an element divisible byf, which is never an optimal embedding. Last, the sublattice is enumerated over a bounded range of majorant norm.Nothing beyond that range is reached, so
Embedcan still report that it found no element where one exists, and it then says which range it examined and whether the range or its own budget ended the search, or which range it would have had to examine when that lies beyond what it is implemented for.Two regimes do this, and they want different things. At a large Eichler level with a small reduced norm it is the range that is exhausted, and widening it does not help: multiplied by
10^6the search spends its whole budget and still finds nothing, so answering those needs a different method rather than a larger constant. At a small level with a large reduced norm it is the budget that is exhausted, and conjugating an embedding may still answer; for the maximal order of discriminant 6 and reduced norm1000000000002it does. SoAl := "Auto"falls back to that method rather than reporting nothing. It does so whenever the search does not reach an element, whichever of the range, the budget or the size of the problem stopped it; onlyAl := "TernaryForm", which names the method, reports the outcome of the search itself. (Issue #59, Issue #48)
Representation Theory
For \(p \equiv 3 \pmod 4\) and ramified \(E/\Q_p\),
AdmissiblePairreturned the wrong \(\chi\) in the pair \((E,\chi)\): on \(\Z_p^*\) it lacked the quadratic character of \(E/\Q_p\). SoGaloisRepresentationandWeilRepresentationof such a local component returned a representation with the wrong determinant, ramified even for a newform of trivial character. Above conductor \(p^3\) its root number could also be wrong. Built from the corrected \(\chi\), the representation now has the right determinant, and its field and Galois group can be larger than before. For ramified \(E\) at any odd \(p\), the two now answer only at conductor \(p^3\) for a newform of trivial character, and raise a not-implemented error otherwise.Inductionof a Galois representation \(A\) over a \(p\)-adic field \(K\) to a subfield \(K_0\) worked only whenField(A)was Galois over \(K_0\). Otherwise, it reported that it could not embedField(A)into the full Galois field, even for the trivial character. It now works without that hypothesis.InductionandPermutationCharactercould also fail on a tower containing a layer of degree one, such as the layerUnramifiedExtension(K,1)adds. The errors were “F is not given as an extension of K” or “Could not find a primitive element”. Such towers are now accepted.RepresentationDimensionraised “Could not decompose” on a root datum that does not split into simple factors and a torus on separate coordinates, such as the root datum of \(GL_n\),StandardRootDatum("A", n-1). On a root datumRthat does split but is not semisimple, it raised an error for a weight of lengthRank(R). For one of lengthDimension(R), it could return a wrong dimension unless the simple coroots are standard basis vectors. It now returns the right dimension, reading a weight of lengthRank(R)as weight coordinates and, whenRis not semisimple, one of lengthDimension(R)as a character of the maximal torus.A product of
LieRepresentationDecompositions with an empty factor returned the other factor instead of an empty decomposition. For semisimple root data, the returned weights also had the other factor’s length rather than the product’s. An empty decomposition is the zero module, so such products are now empty.For a root datum
Rthat is not semisimple,WeylDimensionraised an error for a weight of lengthDimension(R). It now accepts such a weight as a character of the maximal torus, and reads weights of lengthRank(R)as before.HighestWeightRepresentationfailed for reductive Lie algebras and groups of Lie type whose root datum was not semisimple.HighestWeightRepresentationandHighestWeightModulealso failed for semisimple root data when the weight was given in the"Standard"basis.These issues have been fixed. Reported by Eran Assaf.
EpsilonFactorandRootNumberof a degree-1 Galois or Artin representation wildly ramified at \(p\) could be off by a root of unity of \(p\)-power order. The value could also depend on the polynomial chosen to define the field. Where they return a value, it is now correct.After
SetVerbose("GalRep", 1)or any higher level,EpsilonFactorandRootNumberof a local Galois representation, such asGaloisRepresentation(E, p)for an elliptic curveE, usually ended in a runtime error about the number of arguments. They now return the same value at every verbosity level, and results computed at level 0 were never affected.EpsilonFactorandRootNumbercould give the wrong sign for a Galois representation with a two-dimensional orthogonal constituent whose determinant is unramified, in two cases. Over a \(p\)-adic field whose different over \(\mathbb{Q}_p\) has odd valuation, such as a ramified quadratic extension of \(\mathbb{Q}_p\) for odd \(p\), the sign was wrong. ThroughGaloisRepresentation, this gave the wrong local root number to some elliptic curves with additive, potentially good reduction at a prime with such a completion, thoughRootNumberof the curve itself was right. When that constituent was not faithful on the representation’s Galois group, the sign could also be \(-1\) where it is \(1\), over any \(p\)-adic field and for Artin representations. Both cases now give the right sign, but over a field ramified over \(\mathbb{Q}_p\),EpsilonFactorof these representations still has the wrong absolute value. (Issue #399)EpsilonFactorandRootNumbercould be wrong for a Galois representation over a ramified extension of \(\mathbb{Q}_p\) whose residue field is larger than \(\mathbb{F}_p\), including those thatGaloisRepresentationreturns for elliptic curves over such a field. A ramified character could get epsilon factor 0, so thatRootNumberraised a division-by-zero error, and other values could be wrong without an error. This has been fixed. Over any ramified field,EpsilonFactorof an irreducible two-dimensional orthogonal representation with unramified determinant is still too large by a factor \(q^{d/2}\), where \(q\) is the size of the residue field and \(d\) the exponent of the different of the field over \(\mathbb{Q}_p\); this does not affectRootNumber.Over a \(p\)-adic field ramified over \(\mathbb{Q}_p\),
EpsilonFactorcould return a wrong value for an unramified Galois representation of dimension other than 1, and returned 0 for some representations induced from a character of an unramified extension. Whenever the value was 0,RootNumberraised a division-by-zero error.EpsilonFactornow returns the correct value for the unramified representations and no longer returns 0 for the induced ones.
Riemann Surfaces
For a point \(P\) and a divisor \(D\) on a Riemann surface,
P + DandP - Dfailed with a “Bad argument types” error. Both now return a divisor on the same surface, andP + DequalsD + P.A divisor \(E\) whose
Supportis empty, such asZeroDivisor(X)orD - D, made three operations fail an assertion:k*Efor a nonzero integer \(k\),D - Efor any divisor \(D\), and the sum of two such divisors. They now return the zero divisor, or \(D\) in the case ofD - E. Reported by Eran Assaf.
Simplicial Homology
LensSpacenow raises an error when \(p \le 0\), where it silently returned a complex that is not \(L(p,1)\). It also rejects \(p > 23167\), where it ran out of memory or never finished; below that bound a large \(p\) is still accepted and can run for hours.
Strings
WordWraprefused a line width of 1, reporting that its second argument must be a positive integer, which 1 already is. A width of 1 now works and puts one character on each line. A width of 0 or less is still refused, with the same message.The string utilities for the type
MonStgEltwere undocumented. The chapter “Character Strings” now documents them all, among themJoin,Trim,WordWrap, andStringToRational, with a worked example.